The British Columbia Ministry of Environment mapped the inventory of 638 shallow landslides for the winter of 2006–2007 (Fig. 1). These landslides were mapped from 5 m × 5 m resolution SPOT satellite imagery using an automated change detection method comparing scenes from the summer (May to September) of 2006 to the summer of 2007, which is the season when landslides are least frequent (Guthrie et al., 2010b). While the mapped landslides could have occurred anywhere between these dates, it is believed that vast majority occurred during the November 2006 extreme rainstorm (Guthrie et al., 2010b). More detail regarding the mapping and description of these landslides can be found in Guthrie et al. (2010b).

Initiation points digitized from the mapped landslide polygons were used for the subsequent analysis (Fig. 2). The location of the initiation point was manually digitized where the main scarp may be expected. These initiation points were used to approximate the environmental conditions that led to slope failure. Only a single initiation point per landslide was mapped to provide equal treatment of landslides regardless of their size and to reduce the potential effect of spatial autocorrelation in our statistical modelling, which can occur when observations are separated by small distances (Brenning, 2005). The polygons were used to estimate the landslide area. Non-landslide points, which are required for the statistical analysis, were determined by spatial random sampling from the entire study area, excluding the landslide polygons. Overall, 638 observed landslide points and 638 non-landslide points, a 1 : 1 sampling ratio, was used to conduct this study.

The saturation of forest soils from antecedent precipitation has been identified as an important variable for the prediction of landslide initiation (Crozier, 1999; Jakob and Weatherly, 2003). The relationship of precipitation to the landslide inventory was therefore analyzed using the antecedent precipitation leading up to and including the storm on 15 November 2006 (Environment Canada, 2013), which is assumed to have triggered most of the landslides in the inventory (Guthrie et al., 2010b). By observing daily weather station data across the island, it was determined that a 2-week period (2–15 November 2006) would be used to account for these antecedent conditions. These 2 weeks correspond to a period of continuous daily precipitation that preceded the storm.

The actual distribution of rainfall received on the island during the 15 November 2006 storm varied with aspect, elevation, location of storm cells and the wind-driven component of rain. Analysis of meteorological data by Guthrie et al. (2010b) revealed that precipitation varied to more than 200 mm in 24 h. Where rainfall was the only triggering factor, Guthrie et al. (2010b) determined that 88 % of landslides were associated with > 80 mm of rain and 70 % with > 100 mm of rain. In addition, ambient surface temperatures increased by 1–10∘ over the same period and, coupled with strong winds, resulted in considerable snowmelt (a rain-on-snow event) for areas were precipitation alone was not responsible for the landslide response.

The weather station data were compiled from a variety of Canadian government sources: Environment Canada (36 stations), BC Ministry of Transportation (3 stations) and BC Ministry of Forests, Lands and Natural Resources Operations (12 stations). Altogether, 51 stations ranging in elevation from 0 m to 580 m a.s.l. were used for the interpolation of precipitation. Three of these stations were located on small islands that are located within 15 km off the east coast of Vancouver Island. Given these sources of precipitation data, hourly rainfall intensity was not included in our analysis because the majority of our weather station data, which came from Environment Canada, did not have hourly measurements.

On Vancouver Island, the precipitation patterns are strongly dependent on the mountainous topography. Therefore, we applied universal kriging to geostatistically interpolate precipitation based on a spatial linear model that accounts for a relationship to elevation (Goovaerts, 2000). The association between elevation and precipitation is generally dependent on the spatial scale (Daly et al., 2008). In general, an upscaled elevation model can better capture the effects of air movements through topographic obstacles (Daly et al., 1994; Funk and Michaelsen, 2004; Sharples et al., 2005). Therefore, the spatial scale for precipitation interpolation in this study followed the methods used by Daly et al. (2008), which utilized an upscaled 800 m × 800 m resolution DEM.

The performance of the precipitation interpolation was assessed with leave-one-out cross-validation that estimated the root-mean-square error (RMSE) and mean bias (Goovaerts, 2000).

Forest cover was mapped by applying a supervised maximum likelihood classifier (MLC) of Landsat TM scenes from the summer of 2006 reflecting pre-landslide conditions. The MLC was applied to image data obtained by a tasseled cap transformation (Crist et al., 1986). The training data comprised of 100 (visually interpreted) observations per each desired class. The Landsat scenes were first classified and then mosaicked to cover Vancouver Island. The Landsat data were obtained from the United States Geological Survey (USGS) Global Visualization Viewer (USGS, 2011).

The performance of this classification was assessed from reference data collected by visual interpretation of high-resolution DigitalGlobeTM and Spot images from 2003 to 2012 available in Google EarthTM (earth.google.com). The Landsat imagery used for the classification of forest cover was utilized as ancillary data to ensure the reference conditions were relevant to the time of the classification (i.e. the summer of 2006). The locations for the reference data collection were based on the random spatial sample used for the non-landslide point locations. At each reference location a suitable forest cover class was assigned, e.g. sparse, open, semi-open or closed forest. Additionally, an attribute indicating whether the reference sample was in a location that has been previously logged was included. This attribute was used to determine the proportion of each forest cover class associated with logging.

Since forest harvesting has happened and continues to happen on Vancouver Island, the general forest covers were meant to implicitly relate to forest harvesting activities (see class descriptions in Table 1). The classification we used in this study may not be specific enough to only represent “logging forests”. However, the general characteristics of these forest types can be used to relate to forestry practices and investigate the relationships to landslide initiation (Table 1). We based our classification of forest cover on the work of Cohen et al. (1995), who classified different forest stand structures in similar Pacific coast temperate-forest conditions from Landsat TM data.

The classification of forest cover was characterized by comparing the resulting map to BC government forest inventory data. The relative stand density was estimated by crown (canopy) closure values obtained from the vegetation resources inventory (VRI): a geographic data set that represents forest-sampling units with polygons (British Columbia government, 2013a). The VRI represents numerous projected forest conditions for 2013, e.g. stand age, crown closure, dominant forest species. We attempted to represent conditions associated with the year of our study (2006) by including only crown closure estimates in the analysis that had a corresponding projected stand age greater than 6 years.

A BC forestry cutblock inventory (British Columbia government, 2013b) was also compared to our forest classes to gain a sense of how well the classification represents forest-harvesting activities. This inventory consists of digitized cutblock polygons across Vancouver Island for harvesting years up to 2013. The cutblock polygons have a corresponding “end of disturbance date” that was used to approximate the recovery period, in years, associated with our forest cover classes. This value was calculated by subtracting the year corresponding to the “end of disturbance date” to the year our landslides occurred, 2006. Consequently, cutblocks harvested after 2006 were not included in the recovery period calculation. The proportion of samples that were located within the cutblock polygons was calculated to highlight the relative association of each forest cover class to forestry activities. Higher proportions indicate a closer association to the forest cutblocks.

Both comparisons to crown closure and cutblock data were based on the sample of landslides and non-landslide points. Since these data sets do not comprehensively cover Vancouver Island, the comparisons to crown closure and recovery period could only be conducted where the data sets overlapped with the landslide sample. Also, due to the lack of completeness and unknown quality of these government inventories, the comparisons to the forest classification only serve to assess the potential of the forest cover classes as a proxy for recovery stage and were not considered for use as predictors for statistical modelling.

The Euclidean distance from forest service roads was calculated to explore the relationship to landslide initiation in the statistical analysis. Since landslides related to service roads in steep terrain can occur in either the fill material or the cut slopes (Jakob, 2000; VanBuskirk et al., 2005), distances upslope and downslope from roads were considered. Also, assuming road-related landslides were initiated in road fill and road cuts, only distances from roads up to 100 m were considered to have a potential influence on landslides, and consequently the distance variable was trimmed by assigning a value of 100 m to all distance observations > 100 m. Forest road data were obtained from the British Columbia Digital Road Atlas (BCDRA), which was last updated in 2007 (British Columbia Government, 2007). Through visual inspection of Landsat scenes overlaid by the BCDRA, it was determined that the forest service roads were best identified based on road surface attributes for “loose” and “rough” surfaces in the BCDRA data set.

The geologic conditions were represented with generalized formations of rock type, which included igneous (intrusive and volcanic), metamorphic and sedimentary rock. The steep slopes of the Insular Mountains are supported by mainly coarse-grained intrusive rock made of durable minerals (quartz, feldspars) that are relatively resistant to weathering but are subject to mechanical breakdown (Pike et al., 2010). The finer-grained volcanic rocks (e.g. basalts) are highly susceptible to mechanical and chemical weathering. Landslides can occur in volcanic sequences made up of severely weathered rock material or clay residues (Pike et al., 2010). Landslides on Vancouver Island have been observed to occur more likely in metamorphic rocks types such as schist and amphibolite than in granitic intrusive rocks (Guthrie and Evans, 2004a). Also, increased numbers of landslides on the island have been associated with silt and clay sedimentary complexes (Rollerson et al., 2002).

Terrain analysis of topographic characteristics commonly forms the basis of quantitative landslide modelling (van Westen et al., 2008). Slope angle, elevation, plan curvature, profile curvature and upslope contributing area were used to represent topographic conditions related to landslide occurrence. These conditions were derived from a 0.75′′ × 0.75′′ (∼ 20 m × 20 m) resolution Canadian Digital Elevation Data (Canadian Government, 2000) DEM using the standard terrain analysis module implemented in the open-source SAGA GIS (Conrad, 2006).

All of the analysis was performed at a 20 m × 20 m resolution to be consistent with resolution of the DEM that the topographic predictors are derived from. Generally in landslide susceptibility modelling the mapping unit, e.g. pixel size, is based on the DEM resolution (Ayalew and Yamagishi, 2005; Schicker and Moon, 2012). The classified forest cover was resampled from 30 to 20 m spatial resolution using the nearest neighbour method. Proximity to forest service roads was calculated at the 20 m pixel size, and the remaining predictors representing geology and precipitation were also resampled to this resolution in spite of their stronger spatial generalization.

The statistical relationship of the potential preparatory conditions to landslide initiation was modelled with the generalized additive model (GAM) with multiple predictor variables. The GAM is a semiparametric extension of generalized linear models, such as logistic regression, that has the flexibility to represent the response's dependence on a predictor variable as either linear or nonlinear (Hastie and Tibshirani, 1990). The advantage of this method over logistic regression is the ability to represent nonlinear effects that are known to exist in many geomorphological analyses (Phillips, 2003; Brenning et al., 2007; Goetz et al., 2011; Muenchow et al., 2012). These supposed nonlinear effects can be modelled with smoothing functions that fit the data without rigid assumptions regarding the dependence on the response (Wood, 2006). A detailed explanation of the numerous smoothing functions can be found in Hastie and Tibshirani (1990) and Wood (2006). In this study, a smoothing function for multiple predictor variables was estimated with penalized thin plate regression splines with the smoothing parameters selected by generalized cross-validation (Wood, 2006). The flexibility of the smoothers was controlled with up to 3 effective degrees of freedom to allow for general modelling of nonlinear terms while also avoiding the potential to overfit.

A binomial GAM can provide a straightforward interpretation of the odds of an event occurring – in our case, a landslide – conditional on different preparatory conditions. The odds of an event that occurs with probability p is p/(1 - p). Odds ratios can be used as a measure of effect size, i.e. the strength of a predictor. The odds ratio is defined as the odds of an event occurring in one type of condition (e.g. open forest cover) divided by the odds of an event occurring in a different type of condition (e.g. closed forest cover). In this study, we utilize the GAM to assess the effect size of the predictor variables while accounting for possible nonlinear relationships.

The association of possible preparatory conditions with landslide initiation was modelled with predictor variables for topography (five quantitative predictors), rock type (one categorical predictor with four classes), forest conditions (one categorical predictor with four classes) and antecedent precipitation accumulation (one quantitative predictor: Table 4). In addition, the interaction between forest cover and precipitation accumulation was explored with a separate model that also included all the other predictor variables and a linear representation of the corresponding interaction term. We included topography, rock type and precipitation accumulation in our models to account for the differences in slope mechanics and hydrology, while focussing on the association with forest harvesting activities using forest cover and proximity to forest service roads. Odds ratios were estimated from the GAM to assess the empirical effect of these predictor variables on landslide initiation. In general, the effect of a variable was considered strong when the odds ratio was greater than 2, which is equivalent to a 2-fold increase of the odds of an event occurring.

Since there is inherently a dependent data structure in spatial data (Brenning, 2005; Atkinson and Massari, 2011), alternative estimates of the predictors' effects size and confidence limits were obtained by applying a non-overlapping spatial block bootstrap to account for possible spatial autocorrelation within the models (Brenning, 2012). The bootstrap is a resampling-based estimation technique that draws independent resamples (with replacement) from the available data set. Following Brenning et al. (2012), spatial dependencies were accounted for by resampling disjoint sub-regions or “blocks” that were obtained by m means clustering (m = 100) of the sample (landslide and non-landslide) coordinates. The bootstrap resampling of m out of m blocks was repeated 1000 times, and odds ratios were obtained from a GAM fitted to each bootstrap data set. The bootstrap results were summarized by the mean and percentile-based 95 % confidence intervals for the odds ratios.

Although predictive modelling was not the main objective of this study, we assess performance by spatial cross-validation estimation to illustrate the general quality of the susceptibility models (Brenning, 2012). This method randomly partitions the samples into k disjoint spatial subsets using k means clustering, trains a model on k - 1 of these subsets, estimates its performance on the remaining subset and repeats this k times until each subset has been used once as a test set. Since the result is dependent on the particular partitioning used, this step is repeated with different random partitions. In this study, a 5-fold spatial cross-validation was repeated 20 times. In each repetition the area under the receiver operating characteristic curve (AUROC; Zweig and Campbell, 1993) and the sensitivity of each model at a specificity of 90 % were estimated; a specificity of 90 % means that (only) 10 % of the non-landslide area is misclassified as susceptible to landslides, creating false-positive predictions. The estimation of sensitivity at this high specificity assesses the ability of a model to spatially predict areas that are highly susceptible to landslide initiation (Brenning, 2012). The differences in model performance estimates were tested using unpaired t tests.

In addition to the statistical modelling, an analysis was conducted to characterize the landslides associated with different forest conditions and the remaining predictors. The density (number of landslides per square kilometre) was calculated for landslides initiated within each forest cover type and within close proximity to forest service roads (< 20 m). The size distribution corresponding to landslide areas for each forest cover type was presented using a density plot based on Gaussian kernel density estimation. Also, the median and interquartile range (IQR) associated with each predictor was calculated for landslide and non-landslide observations. This univariate analysis was simply exploratory and does not account for other preparatory factors as confounders.

All analyses were performed in the open-source statistical software R (R Development Core Team, 2011) using the following packages for spatial data handling, statistical modelling and error estimation: mgcv (Wood, 2006), raster (Hijmans and van Etten, 2012), ROCR (Sing et al., 2005), RSAGA (Brenning, 2008), sperrorest (Brenning, 2012).

Two-week antecedent precipitation was generally higher in the southwest portions of the island and in the Insular Mountains (Fig. 3a). Precipitation strongly correlated with elevation values from the smoothed DEM with Pearson's correlation coefficient of 0.67 which implies that almost one-half of the variation in precipitation was accounted for by elevation. The residual semivariogram of precipitation after accounting for elevation showed a distance of spatial autocorrelation of ∼ 90 km and almost no nugget effect or sub-scale variation (Fig. 4). Universal kriging predictions consequently combine the overall elevation-related precipitation pattern with local interpolation adjustments in closer proximity to weather stations. The uncertainty of the predicted precipitation values was highest at high elevations and for areas of the island that had a low density of weather stations, which was generally in the north (Fig. 3b). Leave-one-out estimation gave an RMSE of 211 mm and a positive bias of 13 mm, which indicates a slight overestimation of mean precipitation and an acceptable precision considering the high precipitation amounts and large spatial variation (median 481 mm, IQR 613 mm).

The forest cover classification had an overall accuracy of 93.5 % (±2.2 % with a 95 % confidence interval) and 0.896 kappa coefficient. The open and closed forest classes had the most confusion between the reference data and the classification prediction (Table 2).

Overall, the forest conditions of 13 998 km2 on Vancouver Island were classified (Fig. 5). The majority (59 %) of the study area was classified as closed forest, followed by semi-open forest (20 %), open forest (15 %) and sparse forest (6 %). As calculated from the reference data, 60 % of the combined proportion of sparse, open and semi-open forest cover samples was in logged areas. In contrast, only 4 % of the closed forest samples were in a location that was logged. The sparse forest samples were least related to logging with a 46 % proportion of samples in logged areas.

The proportion of samples that were located within the forest cutblocks was highest for sparse (33 %) and open forest (31 %) followed by semi-open (16 %) and closed forest (9 %), which indicates that sparse and open forest cover classes had a stronger association to forest cutblocks than the other classes (Table 3; Figs. 5c and 6). Additionally, the recovery period estimates for samples that were within the cutblock polygons were lowest for the sparse (2 years) and open forest (11 years) classes (Table 3; Fig. 6). This result reinforces that the sparse and open forest cover types were more associated to relatively recent forest harvesting disturbances than the other classes. The sparse and open forest cover classes were also associated with the lowest crown closure estimates: 20 and 45 % respectively (Table 3; Figs. 5d and 6). The median estimate of crown closure was the same (55 %) for semi-open and closed forest. However, this may be expected because these classes had the greatest classification confusion (Table 2).

Closed forest, which covers a majority of the island, had the largest number of initiated landslides (46 %), followed by open forest (26 %), semi-open forest (23 %) and sparse forest (5 %; Table 4). The majority of landslides were also initiated in volcanic rock (62 %), followed by intrusive (28 %), sedimentary (7 %) and metamorphic rock (3 %; Table 4). When looking at the number of landslides initiated in each forest cover class per rock type, volcanic, intrusive and sedimentary rock had a similar distribution to the one for the entire study area. That is, for these rock types, 45–48 % of landslides were initiated in closed forest, 18–23 % in semi-open forest, 25–26 % in open forest and 4–9 % in sparse forest. The distribution of initiated landslides per forest class was generally different for metamorphic. It had only 27 % occurring in closed forest, 50 % in semi-forest, 18 % in open forest and 5 % in sparse forest.

Although closed forest had the largest cumulative area affected by landslides (calculated from the polygons representing the entire landslide area; 2.3 km2) and sparse forest the least (0.43 km2), landslides initiated in sparse forest were associated with the largest landslides in terms of area (median = 8490 m2; Table 5; Fig. 7). The smallest landslide area distribution was associated semi-open forest (2981 m2). The size distributions of landslide area was very similar for closed (4734 m2) and open forest (4457 m2).

According to the univariate analysis, the landslide density in open forest conditions (0.079 landsl./km2) was 2.3 times as great as the density in closed forest conditions (0.035 landsl./km2; Table 5). This was the largest difference between of landslide density amongst the forest conditions. Semi-open forest had a landslide density (0.052 landsl./km2) that was almost equal to the average density found for open and closed forest.

The landslide density in areas adjacent to forest roads (< 20 m) was 30.5 times as high as the general density of landslides for the entire study area. However, regarding this and the other univariate-based findings on the landslide characteristics, topographic and other environmental confounders need to be accounted for in order to confirm the results of this exploratory analysis.

In this study we explored the regional-scale empirical effects of forest harvesting related activities on landslide initiation. The combined effect of open or semi-open forest cover, interpreted as regrowth areas, and proximity to forest roads was substantial with a 3–4 times increase in the odds of initiation compared to closed forest conditions and distances greater than 50 m from forest service roads, without considering the interaction between precipitation accumulation and forest cover. When this interaction is accounted for, this combined effect was estimated to increase the odds of initiation by 6–9 times at high precipitation accumulation (1500 mm). Considering that open and semi-open forest types in this landscape can in general be interpreted as stages of forest regrowth after timber harvesting, this result clearly illustrates that conditions related to forest harvesting were associated with an elevated likelihood of landslide initiation during the November 2006 storm.

Modelled landslides in the classified open forest type were found to be twice as a high as the odds of initiation in closed forest. These findings are similar to May (2003), Guthrie (2002) and Turner et al. (2010) where mature forest had the lowest landslide densities. Although this study did not explore detailed causal mechanisms of landslide initiation, differences in forest cover are associated with different soil hydrological and mechanical conditions (Sidle, 1992) that may be responsible for the lower likelihood of initiation in closed forest cover.

In particular, forest canopies may increase stability by reducing the intensity of precipitation reaching the surface (Keim and Skaugset, 2003). Interception has been found to capture 30–50 % of annual precipitation in dense forest canopies (Dingman, 1994). Also, evapotranspiration in temperate climates is typically higher for closed forest with observed rates that are 5–10 times higher than in exposed soil (Jones, 1997). Deforestation can lead to reduced evapotranspiration, an increase of the amount and intensity precipitation reaching the soil and, as a result, a reduction in the time it takes for the groundwater table to rise, which can lead to slope failure (Johnson et al., 2000). While the protective effect of vegetation may be less pronounced during long-lasting high-intensity precipitation events as observed in this study, we suggest that these effects contribute to both the overall empirical effect of forest cover type and the interaction of forest cover with precipitation in this study.

It was also found that the open forest conditions, representative of young forest, were relatively more susceptible to landslide initiation than sparse forest, representative of recently disturbed forest. This result may empirically support previous findings that there is a lag period between logging activities and increased proneness to landslides (Swanston and Swanson, 1976; Wu et al., 1979; Sidle et al., 2006). Forest harvesting can increase the frequency of landslides during heavy rainfall, with younger forests having more landslides triggered by rainfall events with relatively lower return periods than more developed forests (Imaizumi and Sidle, 2012). A possible explanation for the mechanism involved in this lag period is the deterioration of tree roots, which can increase susceptibility of landslide initiation during high intensity rainfall events (Sidle, 1992).

For the majority of reference data sampled, logging was the most important contributor to the classified forest cover conditions. It was generally observed during the collection of reference data that other site conditions that lead to sparse or open forest were exposed bedrock, thin patches of forest located in high mountain regions and landslide run-out or snow avalanche paths. Thus, although some of the landslides initiated in sparse forest may also be related to timber harvesting activities, other forest disturbances may also play a role in these areas (Guthrie and Evans, 2004b). However, relative to logging disturbances of forest stands in the BC coastal temperate forests, natural disturbances observed at a regional scale are only of a small extent (Pearson, 2010).

Areas adjacent to forest service roads were also found to increase landslide susceptibility. Prior to the November 2006 rainstorm, it was already observed that there were higher densities of landside initiation near (< 20 m) forest service roads for several catchments across Vancouver Island (Guthrie, 2002). Our results suggest that the effect of forest roads on landslides, generalized across Vancouver Island, is up to 50 m from the road. This distance is relatively short compared to the findings of Larsen and Parks (1997) in humid tropical mountains of Puerto Rico and to those of Brenning et al. (2015) in humid Andes of southern Ecuador, who observed the effects of highways on landslides up to 100 and 150 m respectively. An earlier study conducted in the Oregon Coast Ranges, also located in the Pacific Northwest, found most of the road-related landslides initiated during severe winter storms occurred in fill material (Mills, 1997). We believe that this may also be true for road-related landslides on Vancouver Island and would support why we observed the relatively short-distance effect on landslides. In another Coastal Western Hemlock area in the Pacific Northwest of BC, it was observed that most road fill landslides were likely due to loss of strength in foundation soils of older roads built before the 1990s (VanBuskirk et al., 2005). Failure on newer roads built after 1995 was usually associated with inadequate road drainage (VanBuskirk et al., 2005). Broadly speaking, out of all of the factors related to landslide initiation explored in this study, the influence of roads is the most controllable from a forestry practice perspective. Adequate management aimed at improving construction and regular maintenance of road drainage structures is critical to mitigate the influence of forest service roads on landslide occurrence (Slaymaker, 2000; Wemple et al., 2001; VanBuskirk et al., 2005).

We based this analysis on a statistical approach, using the GAM to account for possible confounders (topography, rock type and antecedent precipitation) and potential nonlinear relationships. In general, the topographic predictors (slope, elevation, curvatures and upslope contributing area) were strong at accounting for the spatial variation in this analysis of regional landslide initiation. As usual, slope was an outstanding predictor.

Slope angle is a well-established spatial predictor of landslides that is essential for all susceptibility modelling, including physically based models (Lee and Min, 2001; Dai and Lee, 2002; Goetz et al., 2011); in addition, previous studies on Vancouver Island have found it an important variable for statistical susceptibility modelling (Chung et al., 2001; Goetz et al., 2011). The modelling also illustrates the regional importance of topographic curvature for landslide initiation. Chances of landslide occurrence were higher for convergent and concave curvatures. Their combined effect resulted in a 10-fold increase in the odds of landslide initiation. These hillslope depressions can be generally described as geomorphic hollows, which have been associated with repeated landslide occurrence in steep terrain (Sidle and Ochiai, 2006). Johnson et al. (2000) also observed in similar study area conditions, just north of Vancouver Island on Prince of Wales Island, that these hillslope depressions have a higher tendency to slope failure.

Several mechanisms may be responsible for linking elevation and landslide initiation in this study. Elevation was used as a surrogate for climatic differences and the associated biogeographic characteristics, such as tree species distribution. We observed that the likelihood of initiation gradually decreased in the subalpine and alpine regions (> 800 m a.s.l.). This climate region has soils that are generally saturated throughout the year with most of the precipitation (20–70 %) falling as snow (Meidinger and Pojar, 1991). We conjecture that the reduced likelihood of landslide initiation may be attributed to the presence of a snowpack developed before the 15 November storm. Snowfall is usually the dominant form of precipitation over 800 m a.s.l. on Vancouver Island (Gutrhie et al., 2010b). However, the role of snowpack and snowmelt on landslide occurrence is not very well known; a better understanding of this relationship may help to improve prediction of landslides (Guthrie et al., 2010b).

A previous study on Vancouver Island by Guthrie and Evans (2004a) found that landslides were 1–4 times more likely to occur in metamorphic complexes containing schists and amphibolites. However, they acknowledged that it was difficult to know the role logging activities had on landslide distribution. In our study, these affects and others (e.g. topography) were accounted for in our GAM models. We similarly observed that the effect of rock type on landslide susceptibility was greatest for metamorphic rock types. Our regional estimate observed that landslides were twice as likely to be initiated in metamorphic rocks than in volcanics, the dominant rock type on the island. Guthrie and Evans (2004a) suggest that the degree of metamorphism associated with complexes containing schists and amphibolites has increased the vulnerability to erosion.

It is important to note that we explored the regional characteristics of landslide initiation on Vancouver Island. Thus, the results obtained represent average conditions and characteristics of landslides that occurred during the high-intensity rainfall event. Specific site characteristics of landslide initiation are generally different for catchments across Vancouver Island (Guthrie, 2002). Therefore, we expect that modelled effects on landslide initiation would also vary depending on specific catchment characteristics. The regional modelling approach provides an overview of the underlying effects on landslide initiation to learn more about the corresponding controlling processes, which in turn can be applied to improve model predictions for regional landslide susceptibility.

The landside sampling design may pose some challenges for accurately modelling initiation conditions. The spatial prediction of landslide initiation was modelled based on a single pixel per landslide, ∼ 400 m2 area, that was mapped in the main scarp; however, initiation of debris flows may also occur in non-head scarp regions. Ideally, the exact point of initiation would be modelled. Additionally, only the pixel associated with initiation was labelled with a particular forest cover type. The surrounding or coarser-scale forest cover conditions were not considered. We observed that many of the landslides in the closed forest cover type occurred at or near the boundary between the other forest types. Therefore, we believe that the effect of sparse forest, open and semi-open forest types relative to closed forest may have been underestimated.

The DEM spatial resolution and quality can influence the prediction of landslide susceptibility (Claessens et al., 2005). The quality of the topographic predictors is generally associated with the quality and accuracy of the DEM and the algorithms used to derive them (Raaflaub and Collins, 2006). Previous works on the effects of DEM resolution on prediction of landslide susceptibility found that better model performance was generally associated with DEM resolutions around 10 m (Tarolli and Tarboton, 2006; Penna et al., 2014). For this study we used a 20 m spatial resolution DEM because it was the only available data set with comprehensive coverage of Vancouver Island. However, in terms of predictive ability, we found that the susceptibility models performed well (AUROCs = 86 %) at this resolution for our regional-scale analysis.

The additive models may have underestimated the effect of antecedent precipitation for a 2-week period on landslide initiation as a result of bias introduced by random measurement error (i.e. regression dilution bias; Carroll et al., 2012). Landslide initiation on Vancouver Island is largely related to high-intensity local precipitation (Guthrie and Evans, 2004a). The antecedent precipitation was interpolated based on sparse and poorly located data; the maximum elevation of the weather stations was only 580 m a.s.l. The RMSE and kriging variance provide an indication of how much local variation we are missing in our estimate of antecedent precipitation distribution. The “true” effect in our additive models would be expected to be larger than the estimated one. The thematic coarseness associated with our representation of geology by rock type may have also contributed to regression dilution.